Abstract:
Abstract
Question answering extracts precise answers of a given query (question) to fulfill user‟s information need. However, question answering has limitations in supporting different users such as visually impaired, aged, and illiterate people. Besides, people prefer mobile devices to access information. Formulating queries on small keyboards of mobile devices is cumbersome and tedious. Therefore, to address these problems speech based question answering is needed. This study explored the possibility of developing a prototype Amharic speech based question answering system for open domain factoid questions. Towards this end, literatures are reviewed and tools such as Sphinx tool for speech recognition, Lucene tool for question answering and Netbeans to integrate speech recognition and question answering are used. The Amharic speech-based question answering system accepts speech queries (question) from the user and converts to text using the Amharic speech recognition subsystem. The Amharic question answering subsystem takes the text output of the Amharic Speech Recognition to extract the correct answer from the corpus. Finally, ranked answers are displayed in the form of text. The experimental result shows that the performance of continuous Amharic speech recognition developed for question corpus registered 4.5% WER using development testing and 84.93% recognition performance using live speech input data. The performance of question answering is 76% average precision in retrieving correct answers. After integrating the speech recognition and question answering, the performance registered using speech based question answering system is 75% average precision in retrieving correct answer of a given question. Even though we have achieved noticeable result, the speech recognition system is highly affected by word alignment error and quality of training dataset. Therefore, building standardized Amharic speech corpus and developing good Amharic word aligner is needed to enhance the performance of Amharic speech based question answering.